Академический Документы
Профессиональный Документы
Культура Документы
net/publication/334492340
CITATION READS
1 158
4 authors, including:
Jawad Thajeel
University of Nantes
7 PUBLICATIONS 6 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Effect of spatial variability of soil and rock properties on the bearing capacity of footings View project
All content following this page was uploaded by Abdul-Hamid Soubra on 22 July 2019.
By
Abstract: This paper aims at computing the probability of failure of strip footings resting on
a spatially varying soil and subjected to a vertical load. The active learning reliability method (called
AK-IS) which is a combination of Kriging metamodeling and importance sampling (IS) is used. The
AK-IS technique significantly reduces the computation time with respect to the classical active
learning reliability technique (called AK-MCS) combining kriging with Monte Carlo Simulations
(MCS) by sampling around the design point. It was shown that the critical realization corresponding
to the design point exhibits a perfect symmetry about the central vertical axis of the foundation.
Keywords: strip footing, bearing capacity, spatial variability, kriging, importance sampling,
probability of failure.
I. Introduction
in literature using the crude Monte Carlo Simulations (MCS) or a variance reduction technique (e.g
[1-11]). Despite of being robust and accurate, MCS shows a low efficiency when considering
practical problems with small Pf values especially if a small value of the coefficient of variation on
this failure probability is desired. This is due to the large population required in this case thus leading
to a significant number of evaluations of the performance function. For instance, one million of
samples are required for the computation of Pf values in the order of 10-4 for a coefficient of
1
variation on Pf of 10%. Furthermore, when dealing with spatially varying soil properties as is the
case in the present paper, the evaluation of the performance function is generally based on
computationally expensive finite element/finite difference codes. This naturally leads to a high
computational cost. The variance reduction techniques such as subset simulation (SS) or asymptotic
sampling (AS) significantly reduce the required number of evaluations of the performance function
with respect to the crude MCS; however, these methods remain quite expensive for the computation
Recently, several metamodeling techniques have been developed for the probabilistic analysis
of engineering systems such as the polynomial chaos expansion and its extension the sparse
polynomial chaos expansion, the artificial neural networks, the support vector machine and the
kriging. These techniques have shown high efficiency when the user is interested in the computation
of the first two statistical moments (i.e. the mean and the standard deviation) of the system response
[e.g. 12-15]. Notice however that for problems involving the computation of small failure
probabilities, a large set of sample points is required to accurately construct the meta-model in the
zone of interest for the computation of the failure probability (i.e. the tail distribution). This task is
In order to overcome the shortcoming of the above-mentioned methods related to the large
number of calls to the mechanical model, a combined use of a metamodeling technique with a
simulation-based method (e.g. Monte Carlo, importance sampling, subset simulation) was proposed
by several authors [cf. 16-22]. Among these methods, a combined use of a kriging metamodeling
technique with a Monte Carlo Simulation (MCS) methodology was suggested by Echard et al. [18].
This method is an active learning reliability method combining kriging and Monte Carlo simulation
(called AK-MCS). It overcomes the shortcomings of the crude MCS and the kriging metamodeling
technique when used separately. This method consists in constructing a meta-model (i.e. an
2
analytical equation which substitutes the original mechanical model) based on a relatively small
number of calls to the computationally expensive mechanical model. The computation of the failure
probability may thus be easily performed using this meta-model. It should be emphasized here that
AK-MCS makes use of a powerful learning function (based on the kriging mean prediction and the
kriging variance prediction) for the selection of the ‘best’ samples to be evaluated by the
Echard et al. [18] have illustrated the efficiency of the AK-MCS method through the
computation of the failure probability for some academic examples for which the system response is
known analytically (i.e. where the computation time of the corresponding performance function is
quasi-negligible). Later on, [23] used the AK-MCS technique by Echard et al. [18] for the
computation of the failure probability against soil punching of a strip footing resting on a spatially
varying soil. This problem required the use of a computationally expensive mechanical model based
on numerical simulations for the computation of the performance function. A much reduced number
of calls to the mechanical model was obtained when using the AK-MCS method as compared to the
Although AK-MCS significantly reduces the computation time with respect to the variance
reduction techniques and the meta-modeling techniques, the computation time of this method
remains important. This is because the kriging predictions (mean prediction and variance prediction)
via the meta-model should be evaluated for the whole Monte Carlo population each time a ‘best’ new
sample (called hereafter added sample) is to be selected for evaluation by the mechanical model.
This makes the AK-MCS time-consuming especially when dealing with the small practical values of
the failure probability. This statement was also reported by Echard et al. [18]. In order to overcome
the inconvenience related to the huge number of predictions by the meta-model, this paper makes use
of the active learning method combining kriging with importance sampling IS (called AK-IS
procedure) suggested by Echard et al. [19]. The aim is to perform a probabilistic analysis of the same
3
problem considered by [23] with a more powerful probabilistic technique. In the framework of this
approach, the small failure probability can be estimated with a similar accuracy as AK-MCS but
using a much smaller size of the population (i.e. a much smaller number of calls to the kriging meta-
model each time a new sample is to be selected for evaluation by the mechanical model) because the
sampling population is centered at the design point. This reduced number of calls to the kriging
meta-model naturally leads to a reduction in the computation time with respect to AK-MCS approach
especially for the very small values of the failure probability that require a significant number of
added samples.
Contrarily to [19] where the determination of the design point for importance sampling
computation is straightforward (because the performance function used by these authors was given
by an analytical equation), the computation of the design point becomes an issue in the present case
of spatially varying soil properties where an analytically-unknown performance function with several
random variables is involved in the analysis. This paper presents a simple and non-expensive
iterative procedure based on kriging metamodeling for the determination of the design point in the
present case of a spatially varying soil medium characterized by a quite large number of random
variables. This is followed by the enrichment process to lead to a sufficiently accurate meta-model
The soil cohesion and angle of internal friction were considered as random fields. The
Expansion Optimal Linear Estimation (EOLE) methodology was used to generate these two random
fields. As mentioned above, the mechanical model used in the probabilistic analysis was the one
presented in [23]. It is based on numerical simulations using the finite difference code FLAC3D. The
same deterministic and uncertain parameters considered in [23] were also conserved in this paper for
comparison purposes.
The paper is organized as follows: The next two sections aim at presenting EOLE methodology
and the proposed AK-IS procedure in the case of geotechnical structures involving spatially varying
4
soil properties. This is followed by the probabilistic results. After a validation of the present AK-IS
approach via a simple academic example, some probabilistic results involving a strip footing resting
on a spatially varying soil are presented and discussed. The paper ends with a conclusion of the main
findings.
The Expansion Optimal Linear Estimation (EOLE) method by [24] is used herein to discretize the
two random fields of c and φ. The present two random fields are denoted by Z iNG (x , y ) ( i c , ).
They are described by two non-Gaussian marginal cumulative density functions Gi ( i c , ) and a
common square exponential autocorrelation function ZNG [(x, y), (x', y')] as follows:
x x ' 2 y y ' 2
NG
[( x, y), ( x ', y ')] exp (1)
Z
ax a y
In the present discretization method, one should first define a stochastic grid composed of N q
equation (1) should be transformed into the Gaussian space using the Nataf correction functions
proposed by [25]. As a result, one obtains two Gaussian autocorrelation matrices c ; and ; that
can be used to discretize the two Gaussian random fields at any point using the following equations:
ji
. ji
M
Z i (x , y ) µi i
T
.iZ ( x , y ); i c, (2)
j 1 i
j
where µi and i ( i c , ) are respectively the mean and standard deviation values of the two
random fields, ij ( i c , ; j=1, …, M) are two blocks of independent standard normal random
variables, ij , ij ( i c , ; j=1, …, M) are the eigenvalues and eigenvectors of the two Gaussian
autocorrelation matrices c ; and ; respectively, iZ ( x , y ); is the correlation vector between the
5
values of the random field at the different nodes and its value at an arbitrary point (x, y) as obtained
using Equation (1), and finally M is the number of terms (expansion order) retained in the EOLE
method. This number will be determined later in this section based on the variance of the error. Once
the two Gaussian random fields (i.e. equation 2) are obtained, they should be transformed to the non-
Z iNG (x , y ) G i1 Z i (x , y ) i c, (3)
where (.) is the standard normal cumulative density function. It should be mentioned here that the
series given by Equation (2) are truncated for a number of terms M (expansion order) smaller than
the number of grid points N q , after sorting the eigenvalues cj and j (j=1, …, N q ) in a
descending order. This number should assure that the variance of the error is smaller than a
prescribed tolerance. Notice that the variance of the error for EOLE for a given number s of terms is
2
s 2
1
Var Z i ( x, y ) Z i ( x, y ) Z 1 i
T
i
iZ ( x , y ); (i c, ) (4)
j 1 j
j
where Zi ( x, y) and Z i (x , y ) are respectively the exact and the approximate values of the random
fields at a given point (x, y). In this paper, a maximal value of 5% was adopted for the variance of
the error when discretizing the two random fields (see column 3 of Table 5).
This paper aims at extending the AK-IS approach by [19] to the case of a spatially varying soil
where the computationally expensive mechanical model based on FLAC3D software is used in the
analysis. Details on kriging metamodeling were not provided herein and the reader may refer to [26]
or to different recently published papers where kriging metamodeling is used as in [18-19] and [23].
Also, the details on AK-IS as presented by [19] is not provided herein in order to avoid repetition.
6
Only its extension to the case of spatially varying soil properties was presented in some details in this
paper. It should be mentioned here that the random response predicted by a kriging surrogate model
is a Gaussian variate G N G , G2 where G and G2 are the mean prediction and the
corresponding mean square error (kriging variance) respectively. The variances of the training
samples are zero, but the variances of the other samples are always different from zero.
The present AK-IS procedure consists of two main stages. First, the most probable failure point
(design point) is determined using an approximate kriging meta-model based on a small number of
samples. Second, the obtained approximate kriging meta-model is successively improved via an
enrichment process (by adding each time a new sample selected from a probability density function
hx (X ) centered at the design point) until reaching a sufficiently accurate meta-model for the
computation of the failure probability. These two stages are described in more details in the next two
subsections.
When dealing with problems that are characterized by an explicit performance function, the
design point may be easily determined by minimizing the Hasofer-Lind reliability index subjected to
the constraint that the performance function equal to zero (see [19]). Notice however that when
dealing with analytically-unknown performance functions with several random variables (as is the
case in the present work where spatially varying soil properties are involved in the analysis), the
determination of the design point is less straightforward. The problem is even more difficult when a
high-dimensional stochastic problem is involved (cf. [27]). Indeed, the discretization of the two
random fields of c and φ leads to a significant number of standard normal random variables (between
6 and 62 random variables) as it will be shown later in this paper. The large number of random
variables requires a significant number of calls to the mechanical model for the determination of the
design point.
7
In order to determine the design point in the present work using a relatively small number of
calls to the mechanical model, an iterative procedure based on kriging metamodeling was proposed.
This procedure may be described as follows (see also the flowchart presented in Figure 1):
number of random variables adopted in EOLE methodology for the discretization of both c and
. It should be emphasized here that each sample of M standard Gaussian random variables
provides (when substituted into Equations 2 and 3) typical spatial variations of c and that
respect the correlation structure of these fields, i.e. the so-called ‘realizations’ of c and . The
difference between the different realizations lies in the position of the weak and strong soil
zones although all realizations respect the correlation structure of the corresponding random
fields.
2. From the generated population, randomly select a small number of samples (say N1=20 samples)
of M standard Gaussian random variables. Then, use EOLE methodology to transform each
sample into realizations of c and φ that provide the spatial distribution of the soil cohesion and
angle of internal friction respectively. These realizations are obtained through the computation
of the values of c and at the centroids of the different elements of the FLAC3D mesh using
3. Use the software FLAC3D to calculate the performance function value corresponding to each
sample (the performance function used herein is presented later in equation (10) of this paper).
Based on DACE toolbox, construct an initial approximate kriging meta-model in the standard
space using the N1 samples and the corresponding performance function values.
8
4. Find the minimum value of the Hasofer-Lind reliability index and the corresponding design point
by making use of the already-obtained kriging meta-model and by employing the Generalized
Pattern Search (GPS) algorithm within the global optimization toolbox available in Matlab.
5. Generate a small number of samples (5 samples are used in this work) of M standard Gaussian
random variables according to a multivariate standard Gaussian distribution. Then, translate these
samples such that the obtained samples follow a shifted multivariate Gaussian distribution having
a mean vector whose components are equal to the coordinates of the design point in the standard
coordinate system. After the generation of the five samples, transform each sample into
realizations of c and that provide the spatial distribution of the soil cohesion and angle of
internal friction respectively. Finally, for each one of the five samples, compute the
6. Construct a new kriging meta-model in the standard space using all samples of standard Gaussian
random variables generated so far (i.e from step (2) to step (5)).
7. Compute an updated Hasofer-Lind reliability index and its corresponding tentative design point
8. Steps 5 to 7 are repeated several times until the absolute difference between two successive
values of the Hasofer-Lind reliability index becomes smaller than a given tolerance. The required
number of iterations is denoted hereafter as N2. Consequently, the DoE (which is considered in
this paper to represent the number of samples needed to obtain the final design point) is given by
DoE = N1 + 5 × N2.
It should be emphasized that the aforementioned procedure does not intend to accurately determine
the performance function over the entire design space but it focuses on the computation of the design
point using a relatively small number of evaluations of the computationally expensive mechanical
model. Notice that this procedure was not suggested in [23] because one does not need to determine
the design point when dealing with AK-MCS approach. Notice also that the number N1=20 samples
9
used in this procedure was arbitrarily chosen as an initial guess that can be increased if necessary, the
objective being the construction of an initial approximate meta-model that is suitable for the
determination of a first tentative design point using a limited number of calls to the mechanical
model.
Figure 1. Flowchart of the proposed AK-IS procedure (Stage 1: Determination of the design point)
10
III.2. Enrichment process
Further improvement of the already-obtained kriging meta-model is achieved in this stage via
an enrichment process. Referring to Figure (2), the enrichment process can be explained by the
following steps:
obtained design point, M being the number of random variables needed by EOLE methodology
to discretize the two random fields c and . Notice that the samples generated by IS are called
hereafter candidate samples. Among these samples, only a few ones are computed by the
mechanical model; however, all the candidate samples are calculated by the meta-model each
time a new added sample is to be selected for evaluation by the mechanical model, as it will be
shown below. Notice also that the population size N IS is relatively small herein as compared to
the one generated in the AK-MCS procedure by [23] where N Mc =500,000 samples; however,
both populations may lead to relatively close values of the coefficient of variation on Pf as it
may be seen from the numerical results of AK-MCS and AK-IS approaches.
2. Use the DACE toolbox in order to compute (for the whole population containing the N IS
samples) both the kriging predictor values G and their corresponding kriging variance values
G2 using the obtained meta-model. From the obtained values of the kriging predictors G ,
fx X 1 N IS
f X
Pf I F X
hx X
hx X dX
N IS
I X h X
i 1
F i
x i
(5)
x i
11
hx (X ) is the PDF of the shifted multivariate Gaussian distribution and NIS is the number of
samples. Notice that the values of G ( X ) are calculated using the obtained values of the kriging
V ar Pf
COV Pf (6)
Pf
where Var (Pf ) is the variance of the failure probability estimate. It is calculated by the following
equation:
N IS fx Xi
2
1 1
Var Pf
N IS 1 N IS
IF
i 1
Xi
X i
Pf 2
(7)
hx
3. Identify among the whole population of N IS samples, the ‘best’ next candidate sample for
which one will compute the performance function value using FLAC3D. This is performed by
evaluating a learning function U for each sample in the population. The learning function U
G X
U Xi i
i 1, ...., N IS (8)
G X i
The 'best' next sample is the one that has the smallest U value [i.e. min(U)]. It should be noted
here that the ‘best’ chosen sample is the one that mostly improves the limit state surface (G=0)
of the meta-model because min(U) searches for the sample that has a small kriging predictor (i.e.
a sample that is close to the limit state surface) and/or a high kriging variance (i.e. a high
4. If the obtained minimum value of U is smaller than 2, evaluate the performance function value
based on FLAC3D for this ‘best’ candidate and update the DoE by adding the new ‘best’ sample.
Also, re-construct the kriging meta-model again with the updated DoE.
12
5. Repeat the steps 2 to 4 several times until the smallest U value becomes larger than 2. Notice
that the stopping criterion min (U)>2 corresponds to a maximal probability of making a mistake
At this stage, the learning stops and the meta-model is considered sufficiently accurate for the
computation of the failure probability. When the learning stops, one must compute the estimated
values of both the probability of failure Pf and its corresponding coefficient of variation COV Pf
using the obtained kriging meta-model. It should be emphasized here that the 10,000 evaluations of
the learning function U in step 3 were performed for each added sample since the meta-model is
continuously changing during the enrichment process. This number of evaluations is much smaller
than that used in AK-MCS (i.e. 500,000), thus leading to a much smaller computation time (for a
Stopping
condition on learning
min (U)>2
No N Yes
13
III.3. Numerical implementation
software. It includes the random field discretization by EOLE method, the determination of the
design point by an iterative procedure and the construction of a kriging meta-model for the
computation of the failure probability. The implemented Matlab procedure makes several calls to the
FLAC3D code for the computation of the system response (i.e. ultimate bearing capacity on a
spatially varying soil) or the corresponding performance function value for the different soil
realizations. The computation of the system response via FLAC3D software was not presented herein
to avoid repetition and the reader may refer to [12-13] for more details.
Before the presentation of the probabilistic results of a spatially varying soil, it seems
necessary to validate the present AK-IS procedure by comparison of its results with those obtained
by [19] when considering a simple analytical equation. This is the aim of the next subsection.
IV.1. Validation of the present AK-IS procedure via a simple analytical equation
This section focuses on the validation of the present AK-IS procedure through an analytical
G u1 , u 2 0.5 u1 u 2 1.5 u 2 5 3
2 3
(9)
where u 1 and u 2 are two standard normal random variables. A comparison between the results
obtained by the present AK-IS procedure and those provided by [19] was presented in Table 1.
Notice that in [19], the design point was determined using the classical FORM analysis based on the
analytical equation of the performance function. However; in the present AK-IS procedure, this
design point is determined by employing the iterative procedure proposed in the previous section.
The aim is to check and validate the proposed iterative procedure which will be employed hereafter
14
As may be seen from Table 1, the approximate kriging meta-model (which was needed for the
determination of the design point) was constructed using an initial design of experiments of 15
samples and five iterations with 2 samples per iteration. The enrichment process required 4
additional samples. Thus, the total number of samples (or the number of calls to the performance
function) needed in our procedure is equal to 29 samples. This number is close to that needed by the
classical FORM analysis by [19] (i.e. 26 samples) with the advantage that the present approach may
As a conclusion, the iterative procedure proposed in this paper for the computation of the
design point can be considered as a powerful tool and may be used for more complex cases involving
Table 1. Probabilistic outputs and the corresponding number of calls to the performance function Ncalls as
obtained from the two AK-IS methods
This section aims at presenting the impact of the soil spatial variability on the failure probability
against soil punching of a strip footing subjected to a vertical loading. The soil cohesion c and angle
of internal friction φ were modeled as two non-isotropic non-Gaussian random fields. The EOLE
methodology was used to discretize the two random fields. The illustrative statistical parameters of
these two random fields are presented in Table 2. Recall here that the same autocorrelation function
(square exponential) was used for both c and φ. Notice also that the soil dilation angle ψ was
considered to be related to the soil angle of internal friction φ by 2/3. This means that the soil
15
dilation angle was implicitly assumed as a random field that is perfectly correlated to the soil angle
The performance function employed in the analysis is given by the following equation:
qu
G 1 (10)
qs
where qu is the ultimate bearing capacity computed using FLAC3D model making use of the
generated realizations of c and , and qs is the footing applied loading. Concerning the mechanical
model, a strip footing of breadth B=1m that rests on a soil domain of width 13B and depth 5B was
considered in the analysis. As mentioned above, this mechanical model was not provided herein and
Finally, notice that the number N IS of samples used in most subsequent configurations was
equal to 10,000 samples. This number was found to provide (for these configurations) a small value
of the coefficient of variation on the failure probability (<5%) as it will be shown later. The small
size of the sampling population may be explained by the fact that the sampling is performed
according to a probability distribution that is centered at the design point leading to a much larger
IV.2.1. Evolution of the limit state surface during the computational process
As was previously mentioned in this paper, the AK-IS procedure consists of two main stages:
The first stage (called stage 1) consists in computing the design point from an approximate kriging
meta-model constructed using a small number of samples. In the second stage (called stage 2), the
the evolution of the limit state surface with the addition of new samples (or realizations) during the
16
two stages (i.e. stage 1 and stage 2) was investigated (see Figures 3 and 4). A typical case where
ax=10,000m and ay=10,000m was considered in these figures. This configuration was chosen
because it requires only two random variables and thus, the limit state surface can be easily
Figure 3 presents the evolution of the limit state surface with the addition of new samples
during the different iterations of stage 1. Also, Table 3 presents the evolution of the reliability index
HL for the different iterations. This table shows that the accurate value of the reliability index was
obtained from the first iteration in the present case of a homogeneous soil where ax=ay=10,000m.
Notice however that a larger number of iterations was found necessary (N2 is between 3 and 13) for
spatially varying soil mediums as may be seen from the sixth column of Table 5. It should be
remembered here that only the point of the limit state surface which is the closest one to the origin of
the standard coordinates system is expected to be correct at the end of the first stage of AK-IS; the
other points of this limit state being in general not correctly estimated within this stage.
Figure 3. Evolution of the limit state surface with the addition of new samples during the different
iterations of stage 1 when ax=ay=10,000 m
17
Case HL
Initial DoE = 20 samples 2.6205
Initial DoE + 5 samples of iteration 1
2.6345
= 25 samples
Initial DoE + 5 samples of iteration 1
+ 5 samples of iteration 2 2.6340
= 30 samples
Table 3. The evolution of the reliability index for the different iterations
Figure 4 presents the evolution of the limit state surface with the addition of new samples
(from zero to 28 samples) during stage 2; the number 28 being the needed number of added
realizations during the enrichment process. From this figure, one may notice that the limit state
surface is successively improved with the addition of new samples. Notice however that for the last
two iterations, the two curves representing the limit state surface are coinciding. Thus, the limit state
surface cannot be further improved beyond 24 samples. This means that there is no bias in the meta-
Figure 4. Effect of the number of added samples during the enrichment process on the limit state
surface when ax=ay=10,000 m
Table 4 presents the evolution of the probability of failure Pf and its corresponding
This is in conformity with Figure 4 in which no further improvement in the limit state function was
First of all, recall here that the failure probability is computed each time a new sample is added
during the enrichment process. Figure 5 shows the effect of the number of added samples in the
enrichment process on Pf and COV Pf values for a typical case where ax=10m and ay=1m. This
figure also provides the learning function values for the different added samples. The configuration
(ax=10m and ay=1m) was studied because it represents a practical case requiring a significant number
of random variables (32 random variables in the present case as it may be seen from Table 5).
Figure 5 shows that both Pf and COV Pf vary for the small number of added samples. This
is due to the inaccuracy of the kriging meta-model when only a small number of realizations were
considered. Notice however that both Pf and COV Pf tend to converge to a constant value as the
number of added samples increases. It should be mentioned here that 921 samples were needed in the
enrichment process in addition to the DoE before the algorithm stops (i.e. [min(U)]>2). The final
19
-3
x 10
2 0.04
0.035
1.5
COV(Pf)
0.03
Pf
1
0.025
0.5 0.02
0 200 400 600 800 1000 0 200 400 600 800 1000
Number of added realizations Number of added realizations
1.5
Value of U
0.5
0
0 200 400 600 800 1000
Index i of the added realization
As may be seen from Figure 5, the values of Pf and COV Pf reach an asymptote when the
number of added samples is equal to 823. An additional increase in the number of added samples
does not lead to a significant change in the values of Pf and COV Pf . This means that when the
number of added samples becomes equal to 823, the kriging meta-model is accurate enough (i.e. with
no bias) and it can be used to calculate a rigorous value of the failure probability.
Figures (6.a and 6.b) present two typical non-critical realizations of the soil shear strength
parameters corresponding to the safe (G>0) and failure (G<0) domains respectively for the adopted
reference case (i.e. when ax=10m and ay=1m). On the other hand, Figure (6.c) presents the critical
realizations of the soil shear strength parameters corresponding to the obtained design point for the
same configuration.
Contrary to Figures (6.a and 6.b), Figure (6.c) exhibits a symmetrical distribution of the soil
shear strength parameters with respect to the central vertical axis of the foundation. The weaker soil
20
zone is concentrated around the foundation while the stronger soil is far from the foundation. The
weak soil zone under the foundation allows the failure mechanism to easily develop through this
zone thus reflecting the most prone soil to punching. Concerning the non-critical realizations
(corresponding to G>0 or G<0), it can be observed that the realizations corresponding to the safe
domain exhibits high values of the shear strength parameters (cf. Figure 6.a). The high shear strength
parameters resist soil punching and lead to footing safety. On the contrary, smaller values of shear
strength parameters were encountered in the soil mass when dealing with the realizations
corresponding to the failure domain (cf. Figure 6.b). This allows the failure mechanism to easily
In order to better visualize and interpret the distribution of the soil shear strength in the soil
mass, Figure (7) presents the distribution of the soil cohesion and friction angle along a vertical
section (taken at the center of the footing) for the realizations presented in Figure (6). As may be
seen from Figure (7), the non-critical realizations show more fluctuations than the critical realization
corresponding to the design point, with large values in the safe realization and small values in the
realization corresponding to failure. The distribution of the shear strength parameters corresponding
to the critical realization was shown to present fluctuations in the upper part of the soil profile near
the foundation (i.e. in the depth affected by the soil failure mechanism) and tends to be nearly
uniform in the lower part of the soil. One may also see that smaller values of the soil shear strength
parameters were found in the upper part of the soil mass for this critical realization thus allowing the
failure mechanism to easily develop within this zone. Higher values of the soil shear strength
parameters were observed in the lower part of the soil mass far from the foundation, this zone having
21
(a)
(b)
(c)
Figure 6. Typical realizations (a) safe domain, (b) failure domain and (c) design point
22
Figure 7. Vertical cross-section at the footing center
This section aims at presenting the effect of the autocorrelation distances of the random fields
on the probabilistic outputs (i.e. the failure probability and the reliability index).
Figure 8 presents the effect of the isotropic autocorrelation distance (ax=ay) on Pf and HL as
obtained from AK-MCS and AK-IS methodologies. Also, Figures 9 and 10 present the effect of the
autocorrelation distance (ay or ax) on Pf and HL as obtained from the same two methodologies.
Remember here that the AK-MCS results are those provided by Al-Bittar et al. [23]. However, the
AK-IS results are those obtained in the present paper. From Figures 8, 9 and 10, one may observe
that the two methods lead to similar results. The maximal percent difference between the two
23
Figure 8. Effect of the isotropic autocorrelation distance ax=ay on Pf and HL
Figure 10. Effect of the horizontal autocorrelation distance ax on Pf and HL when ay=2 m
The values of the probabilistic outputs obtained by AK-IS approach and corresponding to the
different soil variabilities were given in Table 5. Columns 2 and 3 of Table 5 provide the number of
24
random variables (or the number of eigenmodes) and the corresponding variance of the error of
EOLE methodology for different values of the autocorrelation distances. Columns 4, 5, 6, 7 and 8 of
the same table provide the failure probabilities, the corresponding values of the coefficient of
variation, the size of the DoE (where DoE = N1 + 5 × N2), the number of added realizations and the
total number of calls to the mechanical model [i.e. DoE + Number of added realizations]. Remember
here that a maximal value of 5% was adopted in this paper for the variance of the error of EOLE
methodology. As may be seen from Table 5, the required number of random variables is small for
the very large values of the autocorrelation distances and significantly increases for the small values
Table 5 shows that the number N1=20 samples suggested in the flowchart of Figure 1 was
sufficient for moderate to large values of the autocorrelation distances; however, a higher number of
samples was found necessary when dealing with the two configurations corresponding to the small
values of the autocorrelation distances [i.e. (ax=ay=2m) and (ax=10m, ay=0.5m)] for which a large
number of random variables (about 60 random variables) was required. The greater number of
samples needed for the configurations corresponding to the small values of the autocorrelation
distances may be explained by the likely increasing non-linearity of the limit state surface for these
cases of very heterogeneous soils. It was found that adopting a value of N1 that is equal to the
number of eigenmodes is a suitable choice (to be able to obtain a first tentative design point) for
these configurations.
Table 5 also shows that the number of samples generated around the successive tentative
design points was equal to 5 as suggested in the flowchart of Figure 1. This small number of samples
was found sufficient even for the small values of the autocorrelation distances. This may be
explained by the fact that the initial construction of the approximate metamodel (that is used to
determine the first tentative design point) is the most difficult task when dealing with the
determination of the design point. Once an initial tentative design point was detected, the
25
determination of the subsequent tentative design points becomes quite straightforward. Furthermore,
the number of iterations that is needed to reach the final design point is quite small (between 3 and
13).
Concerning the IS sampling population, the adopted number NIS of samples determines the
coefficient of variation of the computed failure probability. The required number N IS of samples was
determined in this paper for the two following configurations [i.e. (ax=ay=3m) and (ax=10m,
ay=0.8m)] corresponding to moderate values of the autocorrelation distances. This number was based
on a small target value of the coefficient of variation on Pf of about 5%. It was found equal to about
10,000 samples. The number NIS=10,000 samples was then adopted for all the other configurations
corresponding to larger values of the autocorrelation distances (where larger values of the failure
probability are expected). The obtained values of the coefficient of variation for these configurations
were smaller than 5%. This is because for a prescribed number NIS of samples, the coefficient of
variation on the failure probability is smaller for the larger values of the failure probability. Finally,
notice that the two configurations [i.e. (ax=ay=2m) and (ax=10m, ay=0.5m)] corresponding to small
values of the autocorrelation distances and thus to quite small values of the failure probability have
led to large values of the coefficient of variation when adopting NIS=10,000 samples. Thus, an
increase in the number NIS of samples was needed for these cases. A number NIS of 30,000 samples
was adopted for these configurations. The resulting values of the coefficient of variation on the
failure probability were found to be smaller than 5%. The small values of the coefficient of variation
obtained in this paper (smaller than about 5% for all configurations) indicate that AK-IS leads to
accurate results.
Finally, one may observe from Table 5 that the number of added realizations (and the
corresponding total number of calls to the mechanical model) required to lead to a good
approximation of the kriging model seems to be larger for the smaller values of the autocorrelation
distance (because of the likely increasing nonlinearity of the meta-model in the case of highly
26
heterogeneous soils), although there is no regular increase in the number of added realizations with
the decrease in the autocorrelation distance. Indeed, this number depends on the evolution of the
In order to compare the Pf value obtained by the present AK-IS approach to that computed by
the crude MCS methodology, the reader may refer to the crude MCS results provided in Al-Bittar et
al. [23] for the reference case ax=10m and ay=1m. These MCS results were not detailed herein in
order to avoid repetition. Notice that 136,959 calls to the mechanical model were performed while
The values of Pf and COV Pf obtained from the crude MCS are respectively 1.701×10-3
and 6.54%. These values are to be compared with the present AK-IS results [i.e. Pf =1.628×10-3 and
COV Pf =2.99%] and the results obtained by Al-Bittar et al. [23] using AK-MCS approach [i.e.
Pf =1.656×10-3 and COV Pf =3.47%]. The results provided by the three approaches show good
agreement in term of the value of Pf . As a conclusion, the present AK-IS approach gives a quasi-
similar value of Pf as the crude MCS method (considered as a reference methodology for the
probabilistic analysis). Furthermore, AK-IS is more efficient than AK-MCS because of the smaller
sampling population adopted in this approach as compared to AK-MCS method. This leads to a
significant reduction in the computation time during the enrichment process as it will be explained
below.
27
Table 5. Adopted number of random variables and the corresponding value of the variance of error of EOLE
together with the values of 𝑃𝑓 , 𝐶𝑂𝑉(𝑃𝑓 ), size of DoE, number of added realizations and number of calls to the
mechanical model for various soil variabilities
a. Case of an isotropic case (ax=ay)
Number Number of
ax=ay of
Variance
-3 COV Pf Size of DoE Number of
calls to the
of the Pf ×10 = N1 + 5 × N2 added
(m) random % mechanical
error % realizations
variables model
2 62 4.850 0.710 4.64 62+5×9 2128 2235
3 32 4.647 1.718 5.38 20+5×6 1076 1126
5 24 0.953 2.738 2.42 20+5×10 812 882
10 10 0.815 3.404 1.91 20+5×10 243 313
20 8 0.170 3.745 1.91 20+5×3 200 235
50 6 0.016 3.831 1.82 20+5×6 74 124
100 6 0.001 3.933 1.82 20+5×6 90 140
28
The time required by the meta-model (for each added sample) to perform 500,000
evaluations of the learning function U in AK-MCS (where the learning function is based on the mean
prediction and the variance prediction by the meta-model) is more significant than that required to
perform 10,000 evaluations of this learning function in AK-IS as it was mentioned before.
Furthermore, the number of added samples (and the resulting total number of calls to the mechanical
model) in AK-IS is either greater or smaller than that needed in AK-MCS (see Table 6) but it
remains in the same order as the number of added samples needed in AK-MCS except for the
configurations corresponding to the very small values of the failure probability (because of the
greater number of samples close to the limit state surface in the case of AK-IS approach). As a
conclusion, the computation time required by AK-IS during the enrichment process (which is equal
to the number of added samples multiplied by the time required to compute the 10,000 evaluations of
the learning function U by the meta-model) is much smaller than that required by AK-MCS that
makes use of a quite similar number of added samples with a much greater computation time needed
for the 500,000 evaluations of the learning function U by the meta-model. For instance; when
considering the typical case where ax=10m and ay=2m, 12 days (in average) were necessary to
complete the AK-MCS computation, whereas only 3 days were needed in average to perform a
29
Table 6. Number of added realizations and number of calls to the mechanical model as needed by AK-MCS and
AK-IS for various soil variabilities
a. Case of an isotropic case (ax=ay)
AK-MCS AK-IS
ax=ay Number of Number of calls Number of Number of calls to
(m) added to the mechanical added the mechanical
realizations model realizations model
2 742 762 2128 2235
3 995 1015 1076 1126
5 870 890 812 882
10 286 306 243 313
20 210 230 200 235
50 105 125 74 124
100 100 120 90 140
AK-MCS AK-IS
ay (m) Number of Number of calls Number of Number of calls to
added to the mechanical added the mechanical
realizations model realizations model
0.5 427 447 1937 2037
0.8 790 810 1192 1252
1 752 772 921 981
2 672 692 644 689
5 406 426 354 414
10 286 306 243 313
20 190 210 228 273
50 239 259 210 250
100 232 252 194 264
AK-MCS AK-IS
ax (m) Number of Number of calls Number of Number of calls
added to the mechanical added to the mechanical
realizations model realizations model
2 742 762 2128 2235
5 824 844 988 1073
10 672 692 644 689
20 494 514 437 502
50 357 377 313 393
100 256 276 244 324
30
V. Conclusion
The popular active learning reliability method (called AK-IS) by Echard et al. [19] which is a
combination of kriging metamodeling and importance sampling is used in this paper for the
probabilistic analysis of geotechnical structures involving spatially varying soil properties. More
specifically, the probabilistic model developed in this paper aims at computing the probability of
failure against soil punching of a strip footing resting on a spatially varying soil and subjected to a
vertical load. The soil cohesion and angle of internal friction were modeled by two non-isotropic
non-Gaussian random fields that share an identical square exponential autocorrelation function. The
soil cohesion was modelled by a log-normal distribution and the soil angle of internal friction was
modeled by a beta distribution. EOLE methodology was used for the discretization of the two
random fields.
As is well-known, AK-IS approach has the advantages of both kriging (by using the prediction
mean and prediction variance for the determination of the ‘best’ new candidate sample to be
evaluated by the computationally expensive mechanical model) and importance sampling (for the
generation of samples around the most probable failure point). Indeed, contrary to the active learning
method AK-MCS by Echard et al. [18] combining kriging and Monte Carlo simulations (in which
the learning function is computed via the meta-model for the whole Monte Carlo population for each
added point during the enrichment process), the AK-IS method solves this problem by sampling
around the design point using a much smaller size of the sampling population. This significantly
This paper presents a simple and non-expensive iterative procedure based on kriging
metamodeling for the determination of the design point in the present case of spatially varying soil
properties. The other probabilistic procedure related to the enrichment process is quite similar to that
31
The main findings of this study in terms of the obtained numerical probabilistic results can be
summarized as follows:
1. The present AK-IS procedure was shown to be much more efficient than AK-MCS in the
present case of spatially varying soil properties. It provides an accurate value of the failure
probability (i.e. with a small value of the coefficient of variation on this failure probability)
needing a much smaller computation time as compared to AK-MCS. The reduced computation
time results from the fact that the time required by the meta-model (for each added sample) to
perform 500,000 evaluations of the learning function in AK-MCS is more significant than that
required to perform 10,000 evaluations of this learning function in AK-IS. As a conclusion, AK-
IS significantly reduces the computation time compared to AK-MCS. For instance; when
considering the typical case where ax=10m and ay=2m, 12 days (in average) were necessary to
complete the AK-MCS computation, whereas only 3 days were needed in average to perform a
2. The critical realizations at the design point have shown a symmetrical distribution of the soil
shear strength parameters with respect to the central vertical axis of the foundation with a weak
The main findings and the limitation of this study in terms of the developed methodology can be
summarized as follows:
1. The developed procedure related to the determination of the design point was shown to be a
powerful tool since it can handle complex problems involving spatially varying soil properties
2. Similarly to AK-MCS, the AK-IS kriging approach significantly reduces the number of calls to
the mechanical model as compared to the variance reduction techniques usually used in the
32
geotechnical literature in the case of spatially varying soils. Also, AK-IS approach significantly
reduces the computation time related to the number of the predictions by the meta-model as
compared to AK-MCS. Despite these advantages, AK-IS remains insufficient in the case of very
heterogeneous soils [i.e. when (ax=ay) < 2m] because a large number of calls to the mechanical
model (> 2000 calls) is needed for those cases. More advanced probabilistic approaches are
References
[1] Andersen LV, Vahdatirad MJ, Sichani MT, Sørensen JD. Natural frequencies of wind turbines on
monopile foundations in clayey soils — A probabilistic approach. Computers and Geotechnics
2012;43:1–11.
[2] Vahdatirad MJ, Andersen LV, Ibsen LB, Sørensen JD. Stochastic dynamic stiffness of a surface
footing for offshore wind turbines : Implementing a subset simulation method to estimate rare
events. Soil Dynamics and Earthquake Engineering 2014;65:89–101.
[3] Ahmed A, Soubra AH. Probabilistic analysis at the serviceability limit state of two neighboring
strip footings resting on spatially varying soil. Structural Safety 2014;49:2-9.
[4] Yuan J, Papaioannou I, Straub D. Reliability analysis of infinite slopes under random rainfall
events. Geotechnical Safety and Risk V, Edited by Schweckendiek et al., IOS Press, 2015,
doi:10.3233/978-1-61499-580-7-439.
[5] Li DQ, Xiao T, Cao ZJ, Phoon KK, Zhou CB. Efficient and consistent reliability analysis of soil
slope stability using both limit equilibrium analysis and finite element analysis. Applied
Mathematical Modelling 2016;40:5216–29.
[6] Jiang SH, Huang JS. Efficient slope reliability analysis at low-probability levels in spatially
variable soils. Computers and Geotechnics 2016;75:18–27.
[7] Li DQ, Xiao T, Cao ZJ, Zhou CB, Zhang LM. Enhancement of random finite element method in
reliability analysis and risk assessment of soil slopes using Subset Simulation. Landslides
2016;13:293–303.
[8] Xiao T, Li DQ, Cao ZJ, Au SK, Phoon KK. Three-dimensional slope reliability and risk
assessment using auxiliary random finite element method. Computers and Geotechnics
2016;79:146–58.
[9] Huang J, Fenton G, Griffiths DV, Li DQ, Zhou CB. On the efficient estimation of small failure
probability in slopes. Landslides 2017;14:491–8.
33
[10] Jiang S, Huang J, Zhou C. Efficient system reliability analysis of rock slopes based on Subset
simulation. Computers and Geotechnics 2017;82:31–42.
[11] Van Den Eijnden AP, Hicks MA. Efficient subset simulation for evaluating the modes of
improbable slope failure. Computers and Geotechnics 2017;88:267–80.
[12] Al-Bittar T, Soubra AH. Bearing capacity of strip footings on spatially random soils using sparse
polynomial chaos expansion. International Journal for Numerical and Analytical Methods in
Geomechanics 2013;37:2039–60.
[13] Al-Bittar T, Soubra AH. Efficient sparse polynomial chaos expansion methodology for the
probabilistic analysis of computationally-expensive deterministic models. International Journal
for Numerical and Analytical Methods in Geomechanics 2014;38:1211–30.
[14] Al-Bittar T, Soubra AH. Probabilistic analysis of strip footings resting on spatially varying soils
and subjected to vertical or inclined loads. Journal of Geotechnical and Geoenvironmental
Engineering ASCE 2014;140:1–11.
[15] Al-Bittar T, Soubra AH. Bearing capacity of spatially random rock masses obeying Hoek-Brown
failure criterion. Georisk : Assessment and Managment of Risk for Engineered Systems and
Geohazards 2016;11:215–29.
[16] Bourinet J, Deheeger F, Lemaire M. Assessing small failure probabilities by combined subset
simulation and support vector machines. Structural Safety 2011;33:343–53.
[17] Dubourg V, Sudret B, Bourinet JM. Reliability-based design optimization using kriging surrogates
and subset simulation. Structural and Multidisciplinary Optimization 2011;44(5):673-690.
[18] Echard B, Gayton N, Lemaire M. AK-MCS: An active learning reliability method combining
kriging and Monte Carlo Simulation. Structural Safety 2011;33:145–54.
[19] Echard B, Gayton N, Lemaire M, Relun N. A combined importance sampling and kriging
reliability method for small failure probabilities with time-demanding numerical models.
Reliability Engineering and System Safety 2013;111:232–40.
[20] Bourinet JM. Rare-event probability estimation with adaptive support vector regression
surrogates. Reliability Engineering & System Safety 2016; 150, 210-221.
[21] Moustapha M, Sudret B, Bourinet JM, Guillaume B. Quantile based optimization under
uncertainties using adaptive kriging surrogate models. Structural and Multidisciplinary
Optimization 2016;54(6), 1403-1421.
[22] Huang X, Chen J, Zhu H. Assessing small failure probabilities by AK-SS: An active learning
method combining kriging and subset simulation. Structural Safety 2016;59:86–95.
[23] Al-Bittar T, Soubra AH, Thajeel J. Kriging-based reliability analysis of strip footings resting on
spatially varying soils. Journal of Geotechnical and Geoenvironmental Engineering ASCE
2018;144.
34
[24] Li CC, Der kiureghian A. Optimal discretization of random fields. Journal of Engineering
Mechanics ASCE 1993;119:1136–54.
[25] Nataf A. Détermination des distributions de probabilités dont les marges sont données. Comptes-
rendus de l’Académie des Sciences 1962;225:42–3.
[26] Sacks J, Welch WJ, Mitchell TJ, Wynn HP. Design and analysis of computer experiments.
Statistical Science 1989;4:409-23.
[27] Tandjiria V, Teh CI, Low BK. Reliability analysis of laterally loaded piles using response surface
methods. Structural Safety 2000;22:335–55.
35